Accuracy of line probe assays for the diagnosis of pulmonary and multidrugresistant tuberculosis: a systematic review and meta-analysis

Transcription

1 ORIGINAL ARTICLE TUBERCULOSIS Accuracy of line probe assays for the diagnosis of pulmonary and multidrugresistant tuberculosis: a systematic review and meta-analysis Ruvandhi R. Nathavitharana, Patrick G.T. Cudahy, Samuel G. Schumacher, Karen R. Steingart, Madhukar Pai 5 and Claudia M. Denkinger, Affiliations: Division of Infectious Diseases, Beth Israel Deaconess Medical Center, Boston, MA, USA. Division of Infectious Diseases, Yale University School of Medicine, New Haven, CT, USA. FIND, Geneva, Switzerland. Cochrane Infectious Diseases Group, Liverpool School of Tropical Medicine, Liverpool, UK. 5 McGill International TB Centre, McGill University, Montreal, QC, Canada. Correspondence: Ruvandhi Nathavitharana, Division of Infectious Diseases, Beth Israel Deaconess Medical Center, Suite GB, Lowry Building, Francis Street, Boston, MA, USA. Line probe assays have high accuracy for detection of RIF resistance and INH resistance Cite this article as: Nathavitharana RR, Cudahy PGT, Schumacher SG, et al. Accuracy of line probe assays for the diagnosis of pulmonary and multidrug-resistant tuberculosis: a systematic review and meta-analysis. Eur Respir J 7; 9: 675 [ ABSTRACT Only 5% of multidrug-resistant tuberculosis () cases are currently diagnosed. Line probe assays (LPAs) enable rapid drug-susceptibility testing for rifampicin (RIF) and isoniazid (INH) resistance and Mycobacterium tuberculosis detection. Genotype MTBDRplusV was WHO-endorsed in 8 but newer LPAs have since been developed. This systematic review evaluated three LPAs: Hain Genotype MTBDRplusV, MTBDRplusV and Nipro NTM+MDRTB. Study quality was assessed with QUADAS-. Bivariate random-effects metaanalyses were performed for direct and indirect testing. Results for RIF and INH resistance were compared to phenotypic and composite (incorporating sequencing) reference standards. M. tuberculosis detection results were compared to culture. 7 unique studies were included. For RIF resistance ( 5 samples), pooled sensitivity and specificity (with 95% confidence intervals) were 96.7% ( %) and 98.8% ( %). For INH resistance ( 95 samples), pooled sensitivity and specificity were 9.% ( %) and 99.% ( %). Results were similar for direct and indirect testing and across LPAs. Using a composite reference standard, specificity increased marginally. For M. tuberculosis detection (5 samples), pooled sensitivity was 9% ( %) for smear-positive specimens and % (. 7.7%) for smear-negative specimens. In patients with pulmonary TB, LPAs have high sensitivity and specificity for RIF resistance and high specificity and good sensitivity for INH resistance. This meta-analysis provides evidence for policy and practice. This article has supplementary material available from erj.ersjournals.com Received: May 7 6 Accepted after revision: Oct 6 Support statement: This systematic review was commissioned by WHO in preparation for a Guideline Development Group meeting in March 6. CMD and SGS received additional funding from Department for International Development and the Bill and Melinda Gates Foundation. RRN received additional funding through a Scholar Award from the Harvard Center for AIDS Research (NIAID PAI65-) and an Imperial College Global Health Institutional Strategic Support Fund fellowship from the Wellcome Trust. PGTC received additional funding though the National Institute of Allergy and Infectious Disease (NIAID) training grant in investigative infectious diseases (5TAI757-). Funding information for this article has been deposited with the Open Funder Registry. Conflict of interest: None declared. Copyright ERS 7. This ERJ Open article is open access and distributed under the terms of the Creative Commons Attribution Non-Commercial Licence.. Eur Respir J 7; 9: 675

2 Introduction Tuberculosis causes. million cases and.8 million deaths annually and it is estimated that. million cases go undiagnosed each year []. The emergence of multidrug and extensively drug-resistant tuberculosis ( and XDR-TB, respectively) is a major threat to global tuberculosis control []. Culture and drug-susceptibility testing (DST) using solid media can take up to 8 weeks for results [] and faster liquid-based culture techniques still take 6 weeks []. The delays associated with DST lead to prolonged periods of ineffective therapy and ongoing tuberculosis transmission. The development of rapid molecular diagnostic tests for the identification of Mycobacterium tuberculosis and drug resistance has consequently become a research and implementation priority [5]. Line probe assays (LPAs) are rapid molecular diagnostics that can detect M. tuberculosis and drug resistance. Although LPAs are more technically complex (designed for reference or regional laboratory settings) and take longer to perform than the Xpert MTB/RIF assay (Cepheid, Sunnyvale, CA, USA), they have the ability to detect isoniazid (INH) resistance in addition to rifampicin (RIF) resistance unlike Xpert MTB/RIF [6]. LPAs detect RIF and INH resistance by identifying mutations in the rpob, katg, and inha genes. By targeting mutations in the 8-base pair core region of the rpob gene, more than 95% of all RIF resistant strains can be detected [7]. In comparison, the mutations that cause INH resistance are located in several genes and regions [8, 9]. Although mutations in katg and inha account for approximately 8 9% of INH-resistant strains [], an additional 5 % of INH-resistant strains have mutations in the ahpc oxyr intergenic region, often in conjunction with katg mutations outside of codon 5 []. The World Health Organization (WHO) approved LPAs for the diagnosis of M. tuberculosis and RIF resistance in smear-positive tuberculosis in 8 [], guided by a systematic review evaluating two first-generation LPAs: INNO-LiPA Rif.TB assay (Innogenetics, Ghent, Belgium) and Genotype MTBDR assay (Hain Lifescience GmbH, Nehren, Germany) [], both of which assays are no longer used in clinical practice. Newer versions of the LPA technology have been developed [ 7] and additional studies have been published. This systematic review was commissioned by the WHO to guide a policy update on the use of molecular diagnostics. We evaluated the diagnostic accuracy of three LPAs (appendix A in the supplementary material): GenoType MTBDRplus V (subsequently referred to as Hain V ), GenoType MTBDRplus V (subsequently referred to as Hain V ) and Nipro NTM+MDRTB Detection Kit (subsequently referred to as Nipro ), for the detection of RIF and INH resistance and detection of M. tuberculosis. Methods We followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines and methods for systematic reviews and meta-analyses of diagnostic test accuracy [8, 9]. We prepared a protocol for the literature search, article selection, data extraction, assessment of methodological quality and synthesis of results. Search methods We performed a comprehensive search of the following databases (PubMed, EMBASE, BIOSIS, Web of Science, LILACS, Cochrane) for relevant citations (full search strategy reported in appendix C in the supplementary material). Our search was restricted to the time period January to August 5, since the first Hain LPA was introduced in October. In addition, we contacted laboratory experts and the test manufacturers for additional published studies. We also searched reference lists from included studies and previous meta-analyses []. No language restriction was initially applied but at the full-text review stage we restricted studies to English, French and Spanish. Abstracts or conference proceedings were not included. Study selection and data extraction Two review authors (R.R. Nathavitharana and P.G.T. Cudahy) independently assessed titles and abstracts (screen ). Any citation identified by either review author during screen was selected for full-text review. The same two review authors (R.R. Nathavitharana and P.G.T. Cudahy) independently assessed the full-text articles for inclusion (screen ). In screen, any discrepancies were resolved by discussion between the review authors or by arbitration by a third review author (C.M. Denkinger). Two review authors (R.R. Nathavitharana and P.G.T. Cudahy) extracted data from the included studies with a pre-piloted standardised form and crosschecked to ensure accuracy. Disagreement between review authors on data extraction was resolved by discussion or by a third reviewer (C.M. Denkinger). Studies without extractable sensitivity and specificity data were excluded if no further information was acquired after three attempts to contact the study authors. Selection criteria We included cross-sectional, case-control, cohort studies or randomised controlled trials comparing LPAs to a reference standard test (see below), if at least 5 samples were tested. Patients of all age groups with

3 suspected or confirmed pulmonary tuberculosis or were included, regardless of setting or country. Specimen types were limited to sputum. Patients who were already on therapy were excluded from analyses of M. tuberculosis detection (since dead bacilli not detected by culture could be detected by LPAs leading to false positive results) but were included in analyses for RIF and INH resistance detection. The reference standard test for the detection of M. tuberculosis was a positive solid or liquid culture for M. tuberculosis. The reference standard test for the detection of RIF and INH resistance detection was phenotypic DST for our primary analysis for all studies. Where data were available, LPA results were also compared with a composite reference standard, which combined the results from targeted genetic sequencing and phenotypic DST results (see appendix B in the supplementary material for details). Outcome measures Our outcome measures for all questions were sensitivity and specificity. Indeterminate results were excluded from the analyses for determination of sensitivity and specificity and were reported separately (further details in appendix B in the supplementary material). Assessment of methodological quality We used the Quality Assessment of Studies of Diagnostic Accuracy included in Systematic Reviews- (QUADAS-) instrument, a validated tool for diagnostic studies, to assess study quality []. The information needed to answer QUADAS- questions was incorporated in the data extraction sheet. A description of the QUADAS- items and the interpretation in the study context can be found in appendices D and D in the supplementary material. Statistical analysis and data synthesis We performed statistical analyses using STATA (version ; STATA corporation, College Station, TX, USA). The studies were grouped by type of index test and reference standard used. Our QUADAS- analysis was performed using Excel (version.5.; Microsoft, Seattle, WA, USA). Meta-analysis Meta-analysis was performed for each index test if at least four studies were available for the same index test and if there was limited heterogeneity between studies. Bivariate random effects meta-analyses were performed [, ] using the metandi package in STATA for index tests that included enough data to calculate sensitivity and specificity, with 95% confidence intervals. Summary and individual estimates were also presented graphically with the 95% confidence intervals and prediction region. Several studies did not contribute to both sensitivity and specificity but only to one of the two. In order to make complete use of the data for these studies we performed a univariate random effects meta-analysis of the sensitivity and/or specificity estimates separately. Where there were fewer than four studies available or if substantial heterogeneity precluded meta-analysis, a descriptive analysis was performed. Forest plots were visually assessed for heterogeneity among the studies within each index test. Using summary plots, we examined the variability in estimates and the width of the prediction region, with a wider prediction region suggesting more heterogeneity. We anticipated that studies included in the meta-analysis would be fairly heterogeneous and thus sub-groups for analysis were pre-specified as LPA type, specimen type, specimen conditions and smear status. Results Characteristics of included studies From the literature search, we identified 65 citations and reviewed 8 full-text articles. 7 studies were included in this systematic review (figure ) [5, 7, 9]. 6 of these studies contributed data to more than one analysis, resulting in a total of 9 datasets. A list of excluded studies and the reasons for exclusion is presented in appendix E in the supplementary material. Tables and demonstrate the characteristics of the 9 datasets that provided data on RIF and INH (of note, four of these datasets only provided data on RIF but not INH) and the six datasets that provided data on M. tuberculosis detection respectively. The majority of datasets were cross-sectional in design and almost all were performed in either a regional or national reference laboratory setting. 8 datasets evaluated LPA for direct testing on sputum specimens [5, 7,, 5, 7, 9, 5 9,, 6, 7, 9, 5, 5, 57, 59, 6 69, 7, 7, 7, 77 8, 8, 87, 9, 9] and 6 datasets evaluated LPA for indirect testing on culture isolates [7,, 5 8, 5, 7, 9,,, 5, 8, 5 58, 6 6, 68, 7, 7, 7, 75, 76, 78, 79, 8, 85, 86, 88 9, 9, 9]. 8 datasets evaluated Hain V, five datasets evaluated Hain V [5,, 6, 7] and six datasets evaluated Nipro [7, 7, 79]. Very few datasets recorded demographic data or HIV status due to the use of anonymised samples.

4 Potentially relevant citations identified from electronic databases: n=65 Excluded at screen : n= Not relevant based on assessment of title and abstract Full papers retrieved for more detailed evaluation: n=7 Excluded at screen : n= Abstract on poster: n=5 Duplicate data/study: n= Not about pulmonary TB: n= Other LPAs (including MTBDR): n=9 () No primary data: n= Inappropriate ref standard: n=6 No diagnostic accuracy data: n= Unable to translate: n= Excluded for non-extractable data with no responses from authors: n= Included: only head-to-head comparison of target LPAs (in submission at time of search but since published): n= Papers included in the systematic review: n=7 Papers included for rifampicin resistance: n=7 (phenotypic) n= (composite) Papers included for isoniazid resistance: n=7 (phenotypic) n= (composite) Papers included for Mycobacterium tuberculosis detection: n=6 FIGURE Preferred Reporting Items for Systematic Reviews and Meta-Analyses diagram of included studies. TB: tuberculosis; LPA: line probe assay. Methodological quality The methodological quality across all included studies is summarised in figure and presented for each individual study in appendix D in the supplementary material. Many studies did not report all factors that could affect methodological quality. For the patient selection domain, there was unclear risk of bias for 56 out of 9 datasets for RIF and INH resistance and five out of six datasets for M. tuberculosis detection predominantly because the method of sampling of patients was not defined. Applicability concerns were unclear in 8 out of 9 datasets for RIF and INH resistance and one out of six datasets for M. tuberculosis detection that did not specify the type of patients tested or laboratory setting. For the index test domain, there was unclear risk of bias for 66 out of 9 datasets for RIF and INH resistance and two out of six datasets for M. tuberculosis detection because it was not stated whether the person performing the index test was blinded to the results of the reference standard testing. Applicability concerns in this domain were high risk in eight out of 9 datasets that reported variations in test processing that were not according to the manufacturer s recommendations. For the reference test domain, there was unclear risk of bias for many datasets (68 out of 9) for RIF and INH resistance and three out of six datasets for M. tuberculosis detection because it was not stated whether the person performing the reference test was blinded to the results of the index tests. Applicability concerns were low. In the flow and timing domain, the majority of datasets (78 out of 9 and six out of six, respectively) were judged to have a low risk of bias. Indeterminate result and culture contamination rates datasets reported indeterminate results for directly tested specimens with a median of 5.% and range of..5% for rifampicin and 5.6% and.9.5% for isoniazid (appendix F, table S in the

5 5 TABLE Characteristics of included studies for rifampicin (RIF) and isoniazid (INH) resistance detection, grouped alphabetically by index test type First author [ref.] Country (income category) Study design Laboratory setting Population Number tested: RIF Number tested: INH Direct or indirect Smear status Condition of specimen Phenotypic reference standard Hain Genotype MTBDRplus V AL-MUTAIRI [] # Kuwait (A) Case control National reference Culture 5 5 Indirect N/A Unknown Radiometric BacTec 6 ALBERT [] Uganda (B) Cross-sectional National reference At risk for Direct Positive Frozen Non-radiometric ANEK-VORAPONG [5] Thailand (B) Unclear Regional Culture 5 5 Indirect N/A Frozen Non-radiometric ANEK-VORAPONG [5] # Thailand (B) Cross-sectional Regional Smear 6 6 Direct Positive Frozen Non-radiometric ASANTE-POKU [6] # Ghana (B) Cross-sectional Regional Smear Indirect Positive Frozen Proportion method ASENCIOS [7] Peru (B) Cross-sectional National reference Culture Indirect N/A Unknown Proportion method ASENCIOS [7] Peru (B) Unclear National reference Smear Direct Positive Unknown Proportion method AUNG [8] # Myanmar (B) Cross-sectional National reference Smear Indirect Positive Unknown Proportion method AURIN [9] Bangladesh (B) Cross-sectional National reference At risk for Direct Positive Fresh Proportion method BANU [] Bangladesh (B) Cross-sectional Regional At risk for Direct Positive Unknown Proportion method BARNARD [] South Africa (B) Cross-sectional Regional Smear 8 79 Direct Positive Unknown Proportion method BROSSIER [] # France (A) Unclear National reference Unspecified Indirect N/A Frozen Proportion method BWANGA [] Uganda (B) Unclear National reference Unspecified Indirect N/A Unknown Proportion method CABIBBE [5] Uganda (B) Unclear Regional Unspecified 9 9 Indirect N/A Unknown Non-radiometric CABIBBE [5] Uganda (B) Unclear Regional Unspecified 9 9 Direct Both Unknown Non-radiometric CAUSSE [7] Spain (A) Unclear Regional Smear Indirect Positive Unknown Non-radiometric CAUSSE [7] Spain (A) Unclear Regional Smear 8 8 Direct Positive Frozen Non-radiometric CHEN [8] China (B) Cross-sectional Regional Smear 6 6 Direct Positive Frozen Proportion method CHRYSSANTHOU [9] Sweden (A) Cross-sectional Regional Culture Indirect N/A Fresh Non-radiometric CHRYSSANTHOU [9] Sweden (A) Cross-sectional Regional Culture 9 9 Direct Both Fresh Non-radiometric Continued TUBERCULOSIS R.R. NATHAVITHARANA ET AL.

6 6 TABLE Continued First author [ref.] Country (income category) Study design Laboratory setting Population Number tested: RIF Number tested: INH Direct or indirect Smear status Condition of specimen Phenotypic reference standard DAUM [] South Africa (B) Cross-sectional Regional Culture 6 6 Indirect N/A Frozen Non-radiometric DORMAN [] # South Africa (B) Cross-sectional Regional Prior screened Direct Both Fresh Non-radiometric DUBOIS CAUWELAERT [] Madagascar (B) Cross-sectional Regional Culture 5 5 Direct Positive Unknown Proportion method ELISEEV [] Russia (A) Cross-sectional Regional Smear Direct Positive Unknown Non-radiometric EVANS [] South Africa (B) Unclear Regional Culture Indirect N/A Unknown Non-radiometric FABRE [5] #, Multiple (C) Unclear Regional Culture Indirect N/A Frozen Radiometric BacTec 6 FAROOQI [6] # Pakistan (B) Cross-sectional Regional Smear 5 5 Direct Positive Fresh Proportion method FELKEL [7] # Nigeria (B) Cross-sectional Regional Known Direct Unclear Frozen Non-radiometric FERRO [8] Colombia (B) Cross-sectional National reference Unspecified Indirect N/A Frozen Proportion method FRIEDRICH [9] South Africa (B) Cross-sectional Regional Prior screened 9 9 Direct Both Fresh Non-radiometric GAUTHIER [5] Haiti (B) Cross-sectional National reference Smear Direct Positive Unknown Non-radiometric GITTI [5] Greece (A) Cross-sectional Regional Culture Indirect N/A Unknown Proportion method HILLEMANN [5] # Germany (A) Case control National reference Culture 5 5 Indirect N/A Unknown Mixed HILLEMANN [5] # Germany (A) Unclear National reference Smear 7 7 Direct Positive Unknown Mixed HUANG [5] # China (B) Cross-sectional Regional Smear 5 5 Indirect Positive Unknown Proportion method HUANG [5] # Taiwan (B) Unclear Regional Culture 7 7 Indirect N/A Unknown Mixed HUANG [55] # Taiwan (B) Unclear National reference Culture Indirect N/A Unknown Proportion method HUYEN [56] Vietnam (B) Case control Regional Culture Indirect Positive Frozen Proportion method IMPERIALE [57] Argentina (B) Unclear National reference Culture Indirect Frozen Mixed IMPERIALE [57] # Argentina (B) Unclear National reference Smear 7 7 Direct Positive Frozen Mixed Continued TUBERCULOSIS R.R. NATHAVITHARANA ET AL.

7 7 TABLE Continued First author [ref.] Country (income category) Study design Laboratory setting Population Number tested: RIF Number tested: INH Direct or indirect Smear status Condition of specimen Phenotypic reference standard JIN [58] # China (B) Unclear Regional Culture 7 7 Indirect N/A Unknown Absolute concentration KAPATA [59] Zambia (B) Cross-sectional National reference Smear Direct Positive Frozen Proportion method KHADKA [6] Nepal (B) Cross-sectional Unknown Culture 7 7 Indirect N/A Unknown Absolute concentration KUMAR [6] India (B) Unclear National reference Culture Indirect N/A Unknown Non-radiometric LACOMA [6] # Spain (A) Unclear Unknown Culture 6 6 Indirect N/A Frozen Radiometric BacTec 6 LACOMA [6] Spain (A) Unclear Unknown Unspecified 5 5 Direct Both Frozen Radiometric BacTec 6 LI [6] # China (B) Cross-sectional Unknown Smear 7 7 Direct Positive Unknown Proportion method LUETKEMEYER [6] Multiple (B) Cross-sectional Regional HIV Direct Both Fresh Non-radiometric LYU [65] South Korea (B) Cross-sectional Regional Smear 68 Direct Both Unknown Absolute concentration MACEDO [66] Portugal (A) Cross-sectional National reference Smear Direct Positive Frozen Radiometric BacTec 6 MASCHMANN RDE [67] # Brazil (B) Cross-sectional Regional At risk for Direct Positive Fresh Proportion method MIOTTO [68] Italy (A) Cross-sectional Regional Culture 6 6 Indirect N/A Frozen Non-radiometric MIOTTO [68] Italy (A) Cross-sectional Regional Culture Direct Both Unknown Non-radiometric MIOTTO [69] Burkina Faso (B) Cross-sectional National reference At risk for Direct Both Frozen Non-radiometric MIOTTO [69] Burkina Faso (B) Cross-sectional National reference At risk for Direct Both Frozen Non-radiometric MIRONOVA [7] Multiple (C) Cross-sectional National reference Unspecified Indirect N/A Unknown Non-radiometric MIRONOVA [7] Multiple (C) Cross-sectional National reference Unspecified 7 7 Indirect N/A Unknown Proportion method N GUESSAN [7] Cote D Ivoire (B) Cross-sectional National reference Smear Direct Positive Fresh Non-radiometric NATHAVITHARANA [7] # Multiple (C) Case control National reference Culture positive Indirect N/A Frozen Mixed NATHAVITHARANA [7] Multiple (C) Cross-sectional National reference At risk for 55 6 Direct Both Fresh Mixed Continued TUBERCULOSIS R.R. NATHAVITHARANA ET AL.

8 8 TABLE Continued First author [ref.] Country (income category) Study design Laboratory setting Population Number tested: RIF Number tested: INH Direct or indirect Smear status Condition of specimen Phenotypic reference standard NIEHAUS [7] South Africa (B) Cross-sectional Regional Unspecified Indirect N/A Unknown Proportion method NIKOLAYEVSKYY [7] Russia (A) Cross-sectional Regional Smear 6 6 Direct Positive Fresh Mixed NWOFOR [75] Nigeria (B) Cross-sectional National reference Culture Indirect N/A Unknown Proportion method OCHERETINA [76] #, Haiti (B) Unclear National reference Culture 5 Indirect N/A Unknown Non-radiometric RAIZADA [77] India (B) Cross-sectional Regional At risk for Direct Positive Fresh Proportion method RAVEENDRAN [78] India (B) Cross-sectional Regional Culture Indirect N/A Fresh Non-radiometric RAVEENDRAN [78] India (B) Cross-sectional Regional Smear 6 6 Direct Positive Fresh Non-radiometric RIGOUTS [8] Tanzania (B) Cross-sectional National reference Smear Direct Positive Unknown Proportion method RUFAI [8] India (B) Cross-sectional National reference Smear Direct Positive Fresh Non-radiometric SANGSAYUNH [8] Thailand (B) Cross-sectional Regional At risk for 8 9 Direct Both Fresh Proportion method SCHON [8] Sweden (A) Case control Regional Culture 95 Indirect N/A Frozen Absolute Concentration SCOTT [8] South Africa (B) Cross-sectional Unknown All-comers Direct Both Frozen Non-radiometric SHUBLADZE [85] Georgia (B) Cases only National reference Known 6 6 Indirect N/A Unknown Proportion method SIMONS [86] Netherlands (A) Cross-sectional National reference Unspecified Indirect N/A Unknown Non-radiometric SINGHAL [87] India (B) Cross-sectional National reference Smear Direct Positive Unknown Non-radiometric TESSEMA [88] Ethiopia (B) Cross-sectional Regional Smear 6 6 Indirect Positive Unknown Non-radiometric THO [89] Vietnam (B) Case control National reference Culture 5 5 Indirect N/A Frozen Proportion method TOLANI [9] India (B) Cross-sectional Unknown Smear Indirect Positive Unknown Radiometric BacTec 6 TOLANI [9] India (B) Cross-sectional Unknown Smear Indirect Positive Unknown Radiometric BacTec 6 TUKVADZE [9] Georgia (B) Cross-sectional National reference Smear 7 7 Direct Positive Frozen Mixed VIJDEA [9] # Denmark, Lithuania (C) Case control Supra-national reference Culture 5 5 Indirect N/A Unknown Radiometric BacTec 6 YADAV [9] India (B) Cross-sectional Regional At risk for Direct Positive Fresh Proportion method YORDANOVA [9] Bulgaria (B) Cases only National reference Known Indirect N/A Unknown Non-radiometric Continued TUBERCULOSIS R.R. NATHAVITHARANA ET AL.

9 9 TABLE Continued First author [ref.] Country (income category) Study design Laboratory setting Population Number tested: RIF Number tested: INH Direct or indirect Smear status Condition of specimen Phenotypic reference standard Hain Genotype MTBDRplus V BABLISHVILI [] Georgia (B) Cross-sectional National reference Smear 5 5 Direct Positive Fresh Mixed CATANZARO [6] Moldova, India, South Africa (B) Cross-sectional Regional At risk for 9 9 Direct Positive Unknown Non-radiometric CRUDU [5] Moldova (B) Cross-sectional National reference All-comers Direct Both Fresh Proportion method NATHAVITHARANA [7] # Multiple (C) Case control National reference Culture positive Indirect N/A Frozen Mixed NATHAVITHARANA [7] Multiple (C) Cross-sectional National reference At risk for 5 5 Direct Both Fresh Mixed Nipro NTM+MDR Detection Kit MITARAI [7] # Japan (A) Cross-sectional National reference Unspecified Indirect N/A Unknown Unknown MITARAI [7] Japan (A) Cross-sectional National reference Unspecified 55 5 Direct Both Frozen Proportion method RIENTHONG [79] # Thailand (B) Case control Unknown Culture 6 6 Indirect N/A Frozen Non-radiometric RIENTHONG [79] Thailand (B) Cross-sectional Unknown Unspecified 7 7 Direct Both Fresh Non-radiometric NATHAVITHARANA [7] # Multiple (C) Case control National reference Culture positive Indirect N/A Frozen Mixed NATHAVITHARANA [7] Multiple (C) Cross-sectional National reference At risk for 75 7 Direct Both Fresh Mixed : multidrug-resistant tuberculosis. : these studies only contributed data to PICO Aa and b (Rifampicin resistance detection); # : these studies contributed data from which a composite reference standard could be derived. TUBERCULOSIS R.R. NATHAVITHARANA ET AL.

10 TABLE Characteristics of included studies for M. tuberculosis detection, grouped alphabetically by index test type Author, year Country (income category) Study design Laboratory setting Population Number tested Direct or indirect Smear status Condition of specimen Processed per manufacturer s instructions Culture reference standard Hain Genotype MTBDRplus V DORMAN [] South Africa (B) Cross-sectional Regional Prior screened Direct Both Fresh Yes Liquid: MGIT 96 FELKEL [7] Nigeria (B) Cross-sectional Regional Known Direct Unclear Frozen Yes Liquid: MGIT 96 FRIEDRICH [9] South Africa (B) Cross-sectional Regional Prior screened 6 Direct Both Fresh Yes Liquid: MGIT 96 LUETKEMEYER [6] Multiple (B) Cross-sectional Regional HIV 595 Direct Both Fresh Yes Liquid: MGIT 96 SCOTT [8] South Africa (B) Cross-sectional Unknown All-comers 77 Direct Both Frozen Yes Liquid: MGIT 96 Hain Genotype MTBDRplus V CRUDU [5] Moldova (B) Cross-sectional National reference All-comers 6 Direct Both Fresh Yes GenoLyse and GeneXtract Liquid: MGIT 96 : multidrug-resistant tuberculosis. TUBERCULOSIS R.R. NATHAVITHARANA ET AL.

11 a) Patient selection Index test Low-risk of bias High-risk of bias Unclear risk of bias b) Patient selection Index test Reference standard Flow and timing Reference standard Flow and timing QUADAS- domain % of datasets QUADAS- domain % of datasets c) Patient selection Index test Low concern High concern Unclear concern d) Patient selection Index test Reference standard Reference standard QUADAS- domain % of datasets QUADAS- domain % of datasets FIGURE QUADAS- summaries. a and c) Risk of Bias and Applicability Concerns summary about each QUADAS- domain presented as percentages across the 9 included datasets for rifampicin (RIF) and isoniazid (INH) resistance compared with phenotypic culture-based reference standard (of note, four datasets only contributed to RIF). The summaries for the datasets for RIF and INH compared with composite reference standard are not displayed separately since these datasets are a subset of the 9 datasets displayed below and thus the figures displayed are thought to be accordingly representative. b and d) Risk of Bias and Applicability Concerns summary about each QUADAS- domain presented as percentages across the six included datasets for Mycobacterium tuberculosis detection compared with a culture-based reference standard. supplementary material). Only five datasets reported indeterminate results for indirectly tested isolates but these percentages were lower than for direct testing, with a median of.% and range of.5.% for rifampicin and.5% and.5.% for isoniazid. Only three datasets that performed direct testing for M. tuberculosis detection reported indeterminate results, with a median of.% and range of.7.7%. Data on smear grade were limited. Studies did not typically report whether repeat testing was performed on indeterminate results. For comparison purposes, four datasets reported the number of contaminated cultures obtained using the culture reference standard, with a median of 7.6% and range of.8 7.% of the total specimens. Analysis of primary outcomes of interest Diagnosis of RIF resistance using a phenotypic reference standard Pooled Analysis for all LPAs on all specimen types 9 datasets were included in the bivariate analysis, with a total of 5 samples that included 6789 (%) confirmed RIF-resistant tuberculosis cases. Meta-analysis revealed a pooled sensitivity of 96.7% (95% CI %) and specificity of 98.8% (95% CI %) (table ). Results were largely homogenous, with a small proportion of studies being outliers. Pooled analysis stratified by LPA (appendix F, table S in the supplementary material) demonstrated a slightly lower sensitivity for Hain V and Nipro (95.% and 9.% compared with 97.% for Hain V) although confidence intervals overlapped and specificity was similar (98.%, 98.% and 98.9% respectively). Direct testing 8 datasets tested RIF resistance detection with LPA directly from specimens, with a total of 56 samples that included 876 (7%) confirmed RIF-resistant tuberculosis cases. The pooled sensitivity was 96.% (95% CI %) and specificity was 98.% (95% CI %) (table, figure a). Outliers with lower sensitivity and specificity were predominantly datasets with limited numbers of resistant specimens (<) and thus accompanied by very wide confidence intervals (figure ). Indirect testing datasets tested RIF resistance detection with LPA indirectly from isolates, with a total of 696 samples that included 9 (7%) confirmed RIF-resistant tuberculosis cases. The pooled sensitivity was 96.9%

12 TABLE Diagnostic accuracy of line probe assays for all three assays combined for rifampicin (RIF) and isoniazid (INH) resistance and multidrug-resistant tuberculosis () detection Reference standard Test Direct or indirect Smear status Datasets (samples) n Sensitivity (95% CI) Specificity (95% CI) Phenotypic drug susceptibility testing RIF Both All 9 ( 5) 96.7% ( ) 98.8% ( ) RIF Direct All 8 ( 56) 96.% ( ) 98.% ( ) RIF Indirect All ( 696) 96.9% ( ) 99.% ( ) Composite drug susceptibility testing RIF Both All (58) 95.% ( ) 99.5% ( ) Phenotypic drug susceptibility testing RIF Both All (58) 95.% ( ) 98.9% ( ) (same samples as composite drug susceptibility testing) Phenotypic drug susceptibility testing INH Both All 87 ( 95) 9.% ( ) 99.% ( ) INH Direct All 6 ( 7) 89.% ( ) 98.% ( ) INH Indirect All ( 6) 9.% ( ) 99.7% (99. ) Composite drug susceptibility testing INH Both All (56) 85.% ( ) 99.9% ( ) Phenotypic drug susceptibility testing INH Both All (5) 85.% ( ) 99.5% ( ) (same samples as composite drug susceptibility testing) Phenotypic drug susceptibility testing Both All 57 ( ) 9.9% (9. 9.7) 99.% ( ) Composite drug susceptibility testing Both All (75) 86.6% (8.9 9.) 99.6% ( ) Phenotypic drug susceptibility testing (same samples as composite drug susceptibility testing) Both All (75) 86.9% (8. 9.7) 99.5% ( ) (95% CI %) and specificity was 99.% (95% CI %) (table ; and appendix F, figure S8a in the supplementary material). Point estimates for sensitivity for individual studies were even more homogenous than those for direct testing (appendix F, figure S in the supplementary material). The reasons for the outlier studies with lower sensitivities were unclear as the populations tested (all-comers versus those with risk) differed in the respective studies [6, 7, 9]. One outlier demonstrated a lower specificity (78.%, 95% CI %) for specimens tested by solid (Löwenstein Jensen) rather than liquid (Mycobacteria Growth Indicator Tube (MGIT)) culture [7]. Diagnosis of RIF resistance using a composite reference standard Pooled Analysis for all LPAs on all specimen types datasets contained data comparing LPA with a composite reference standard (using the results from targeted sequencing of either the RIF-resistance determining region or rpob gene and phenotypic DST), with a total of 58 samples that included 9 (8%) RIF-resistant M. tuberculosis cases [7,, 6,, 5 7, 5 58, 6, 6, 67, 79, 9]. Most studies only performed sequencing on discrepant results, thus results from this analysis may be potentially biased in favour of the LPAs. Bivariate meta-analysis of these studies revealed a pooled sensitivity of 95.% (95% CI %) and specificity of 99.5% (95% CI %) (table ). Specificity increased when a composite standard was used as 7 LPA false-positive results based on comparison to phenotypic DST (from datasets) were reclassified as true as sequencing confirmed the presence of known resistance-conferring mutations (appendix F, table Sa in the supplementary material). Of note, the sensitivity was lower in this subset of datasets for which data on a composite reference standard could be derived compared with the overall dataset, which we hypothesise may be due to some selection bias in the studies that performed targeted sequencing alongside phenotypic DST. Heterogeneity across studies was limited (appendix F, figure S5 in the supplementary material). MASCHMANN RDE et al. [67] demonstrated a sensitivity of 8.8% and stated that two out of the five specimens incorrectly classified had insertions in codons which may have caused hybridisation of the corresponding wild-type probe (wt for codons 57 5) and the other three were wild-type on sequencing, suggesting that resistance may be driven by mutations outside of the rpob hotspot. Diagnosis of INH resistance using a phenotypic reference standard Pooled analysis for all LPAs on all specimen types 87 datasets were included in the bivariate analysis, with a total of 95 samples that included 85 (9%) confirmed INH-resistant tuberculosis cases. Meta-analysis revealed a pooled sensitivity of 9.% (95% CI %) and specificity of 99.% (95% CI %) (table ). Results were moderately heterogeneous

13 a) b).8.8 Sensitivity.6. Study estimate Sensitivity % confidence region Summary point..8 95% prediction region Specificity.6.. Specificity c) d).8.8 Sensitivity.6. Sensitivity Specificity.6.. Specificity FIGURE Hierarchical summary receiver operating characteristic graphs of summary estimates. Bivariate analysis of the sensitivity and specificity for all line probe assays for the diagnosis of drug resistance detection compared with a phenotypic reference standard for specimens tested directly for a) rifampicin resistance, b) isoniazid resistance, c) multi-drug resistance and d) the detection of Mycobacterium tuberculosis compared to a culture reference standard. In the plots below, the red squares represent the pooled summary estimates, the dashed red lines represent the 95% confidence region and the dashed green lines represent the 95% prediction region. The individual circles represent each study and the size of the circle is proportional to the total sample size. for sensitivity, whereas specificity estimates were more homogeneous. Pooled analysis stratified by LPA (appendix F, table S in the supplementary material) demonstrated a lower sensitivity for Nipro (86.9%) and higher sensitivity for Hain V (9.6%) compared with Hain V (9.%) although specificity was similar (99.%, 99.% and 99.%) respectively. Direct testing 6 datasets tested INH-resistance detection with LPA directly from specimens against a phenotypic reference standard, with a total of 7 samples that included 576 (%) confirmed INH-resistant tuberculosis cases. The pooled sensitivity across studies was 89.% (95% CI %) and specificity was 98.% (95% CI %) (table, figure b). Greater heterogeneity was noted for INH-sensitivity compared with RIF for sensitivity (figure 5). Several outliers had limited numbers of resistant specimens (<) and were thus accompanied by very wide confidence intervals [7,, 8, 8]. Explanations for outlier results included the known geographic variation of mutations and heteroresistance. Indirect testing datasets tested INH resistance detection with LPA indirectly from isolates against a phenotypic reference standard, with a total of 6 samples that included 559 (%) confirmed INH-resistant tuberculosis cases. The pooled sensitivity across studies was 9.% (95% CI %), which was higher than seen with direct testing, as was the case for specificity, which was 99.7% (95% CI 99..%) (table ; and

14 First author (ref.] Sensitivity (95% CI) Specificity (95% CI) TP FP FN TN MTBDRplus V SCOTT [8]. ( ) 96. ( ) 8 RIGOUTS [8] 6. ( ) 98.8 ( ) 6 CHRYSSANTHOU [9] 75. ( ) 98.8 ( ) 85 MASCHMANN RDE [67] 8. (6. 9.9) 9. ( ) 5 FELKEL [7] 8. ( ). (86.77.) 5 6 DORMAN [] 85.7 ( ) 99. ( ) KAPATA [59] 85.7 (. 99.6) 99.6 ( ) 6 57 CHEN [8] 85.9 ( ) 9. ( ) LI [6] 88.9 ( ) 97.7 ( ) 6 5 ASENCIOS [7] 9.67 ( ) ( ) 75 FAROOQI [6] 9.5 ( ) 96.8 ( ) 9 9 SANGSAYUNH [8] 9. ( ) ( ) RAIZADA [77] 9.8 ( ) 9.75 ( ) BANU [] 9.55 (8. 98.) 8.6 ( ) 58 LYU [65] 9.9 ( ). (97.6.) IMPERIALE [57] 96.5 ( ). (9.75.) 5 8 NIKOLAYEVSKYY [7] 96. (9. 99.) 89.7 ( ) 7 TUKVADZE [9] ( ) 98.8 ( ) 8 HILLEMANN [5] ( ). (9.9.) NATHAVITHARANA [7] 97.8 (9. 99.) 97.9 ( ) SINGHAL [87] ( ) 9. (8. 99.) 8 DUBOIS CAUWELAERT [] 97.9 ( ) 98.6 ( ) 7 YADAV [9] ( ) 99. ( ) 7 7 BARNARD [] ( ) 99. ( ) 9 5 CABIBBE [5]. (7.5.). (9..) 5 FRIEDRICH [9]. (.5.). (95.9.) 89 LACOMA [6]. (86.77.) 86.6 ( ) 6 9 LEUTKEMEYER [6]. (8.56.) 95.8 ( ) 8 MIOTTO [68]. (9..) 98.8 ( ) 65 MIOTTO [69]. (8.5.) 8. ( ) 9 MIOTTO [69]. (6.6.) 5. ( ) 8 ALBERT []. (78..) 9.8 ( ) 5 7 ANEK-VORAPONG [5]. (8.5.). (97.9.) 9 5 AURIN [9]. (98.7.). (95.89.) CAUSSE [7]. (66.7.). (66.7.) 9 9 ELISEEV []. (9.79.) 98.8 ( ) 69 GAUTHIER [5]. (76.8.). (98..) 95 MACEDO [66]. (85.8.). (9..) 5 N'GUESSAN [7]. (9.8.) 7.55 (6. 85.) 65 RAVEENDRA [78]. (7.8.). (7.5.) 5 RUFAI [8]. (8.6.). (9..) MTBDRplus V BABLISHVILI [] CRUDU [5] CATANZARO [6] NATHAVITHARANA [7] 9.8 ( ) 9. ( ) ( ) 98.5 ( ) 99.6 ( ) 96. ( ) ( ) 97.8 ( ) Nipro RIENTHONG [79] NATHAVITHARANA [7] MITARAI [7] 75. ( ) 96.9 ( ). (9..). (96.95.) 97.5 ( ). (9.5.) Sensitivity % Specificity % FIGURE Forest plots demonstrating the sensitivity and specificity of all line probe assays for rifampicin resistance-detection for sputum specimens tested directly compared with phenotypic drug susceptibility testing. TP: true positive; FP: false positive; FN: false negative; TN: true negative. appendix F, figure S8b in the supplementary material). Several studies were outliers for sensitivity but specificity was largely homogeneous (appendix F, figure S6 in the supplementary material). Reasons for lower sensitivity include the use of different types of phenotypic DST within a study [7], the presence of less common resistance mutations due to geographic variation and difficulty detecting low-level INH resistance [6]. The outlier for specificity only contained three INH-sensitive strains [9]. Diagnosis of INH resistance using a composite reference standard Pooled analysis for all LPAs on all specimen types datasets contained data comparing LPA with a composite reference standard, with a total of 56 samples that included 6 (5%) INH-resistant M. tuberculosis cases [7,, 5, 6, 8,,, 6, 7, 5 55, 58, 6, 67, 68, 79, 9]. Bivariate meta-analysis of these studies revealed a pooled sensitivity of 85.% (95% CI %) and specificity of 99.9% (95% CI 99.6.%) (table ). Bivariate analysis of the same datasets compared to phenotypic DST revealed a pooled sensitivity of 85.% (95% CI %) and specificity of 99.5% (95% CI %). Sequencing also revealed resistance mutations that were not detected by LPA (appendix F, table Sb in the supplementary material). For example, of the strains with a rarer katg mutation S5N were not detected in the study by JIN et al. [58] due to the lack of the appropriate mutation probe in the Hain V assay and because the wild-type band also failed to disappear. Although seven LPA false-positive results (from six datasets) were reclassified as true in total based on sequencing confirming a known resistance mutation (four katg S5T mutations and three inha c-5t mutations), specificity barely increased when a composite standard was used.

15 First author (ref.] Sensitivity (95% CI) Specificity (95% CI) TP FP FN TN MTBDRplus V RIGOUTS [8] 5.9 ( ) 98.6 ( ) 8 5 MASCHMANN RDE [67] 6. (5.7 7.). (76.8.) 9 9 DORMAN [] 6.7 (.6 79.) ( ) 8 9 SCOTT [8] ( ) 95.8 ( ) 79 RAIZADA [77] 7.89 ( ) 96.8 ( ) 5 6 MIOTTO [68] 7.7 ( ) 9. ( ) 8 5 CABIBBE [5] 7.9 ( ) 9. ( ) 8 7 KAPATA [59] 75. ( ) 97.6 ( ) 8 99 FAROOQI [6] 76.7 ( ). (9.96.) 5 CHEN [8] 76.7 ( ) 95. ( ) 65 LI [6] ( ) ( ) 6 DUBOIS CAUWELAERT [] 79.7 ( ) 97.8 ( ) 55 8 FRIEDRICH [9] 8. ( ). (95.55.) 8 GAUTHIER [] 8.56 ( ) 96.5 ( ) ALBERT [] 8.77 ( ). (9.56.) 5 66 SINGHAL [87] 8. ( ) 9.75 ( ) FELKEL [7] 85.7 (. 99.6). (86.8.) 6 5 LEUTKEMEYER [6] 85.7 ( ) ( ) 57 RAVEENDRAN [78] 85.7 (. 99.6). (66.7.) 6 9 IMPERIALE [57] 87.8 ( ). (85.8.) 5 CHRYSSANTHOU [9] 88. ( ). (95.7.) 5 7 TUKVADZE [9] 89.8 ( ) 99.9 ( ) HILLEMANN [5] 9. ( ). (88..) 7 MIOTTO [69] 9.67 ( ). (59..) 7 YADAV [9] 9.7 ( ) 99. ( ) BANU [] 9.65 ( ) 8. ( ) LYU [65] 9.86 ( ) 98. ( ) 9 SANGSAYUNH [8] 9.86 ( ) 8. ( ) ANEK-VORAPONG [5] 9. ( ). (97..) 7 5 BARNARD [] 9.7 ( ) 99.7 ( ) 7 9 NATHAVITHARANA [7] 9. ( ) 96.9 ( ) 86 9 N GUESSAN [7] 9.9 ( ) 95. ( ) 75 9 LACOMA [6] 96. ( ). (8.89.) 6 ELISEEV [] 96.6 ( ). (95.7.) 86 8 ASENCIOS [7] ( ) ( ) 66 NIKOLAYEVSKYY [7] 97.5 ( ) 79. (6.8 9.) 7 6 AURIN [9] 99.8 ( ) 98.8 ( ) 9 85 CAUSSE [7]. (6.6.). (69.5.) 8 MACEDO [66]. (85.75.). (9.96.) MTBDRplus V BABLISHVILI [] CATANZARO [6] NATHAVITHARANA [7] CRUDU [5] 89.6 ( ) 9.9 ( ) 95. ( ) 96.6 ( ). (98.7.) 99.6 ( ) 98.8 ( ) 86.9 ( ) Nipro MITARAI [7] RIENTHONG [79] NATHAVITHARANA [7] 5. ( ) ( ) 9.9 ( ) 97.8 ( ) 96. (9. 99.) ( ) Sensitivity % Specificity % FIGURE 5 Forest plots demonstrating sensitivity and specificity of all line probe assays for isoniazid-resistance detection for sputum specimens tested directly compared with phenotypic drug susceptibility testing. TP: true positive; FP: false positive; FN: false negative; TN: true negative. Heterogeneity assessment (appendix F, figure S7 in the supplementary material) demonstrated homogenous results for specificity, which was largely also the case for sensitivity aside from a few outliers. MITARAI et al. [7] demonstrated a specificity of 6.6% (95% CI %). Of the 5 isolates incorrectly identified as sensitive by LPA, had a range of rare katg mutations not identified by any of the katg probes, 7 had fabg inha mutations and were identified as wild-type by sequencing. MASCHMANN RDE et al. [67] demonstrated a sensitivity of 6.% (95% CI 5. 7.%) and reported that all 9 strains misclassified as susceptible on LPA were found to have wild-type katg and inha genes according to targeted sequencing, indicating that there may have been mutations in other genes associated with INH resistance or efflux systems that could not be detected by the LPA. Diagnosis of multidrug resistance Pooled analysis for all LPAs on all specimen types 57 datasets included data on the diagnostic accuracy of LPA for detection, with a total of samples that included 8 (%) confirmed cases [ 9,,, 5, 7, 8,, 8, 5, 5 56, 58 6, 66 69, 7 7, 75, 78, 8, 8, 8, 88, 9, 9, 9]. Bivariate meta-analysis of these datasets revealed a pooled sensitivity of 9.9% (95% CI %) and specificity of 99.% (95% CI %)(table and figure c). Figure 6 demonstrates homogenous results for specificity aside from a few outliers in which the number of sensitive (non-mdr strains) was <5, which was largely also the case for these sensitivity outliers. Comparison of diagnostic accuracy from direct versus indirect testing Based on the analysis of all data, the estimates for sensitivity of LPA for RIF and INH resistance were almost identical for LPA performed directly on sputum specimens and indirectly on culture isolates (96.% and 5

16 First author (ref.] Sensitivity (95% CI) Specificity (95% CI) TP FP FN TN MTBDRplus V MASCHMANN RDE [67] RIGOUTS [8] CHEN [8] LI [6] LACOMA [6] JIN [58] HUANG [5] KHADKA [6] DUBOIS CAUWELAERT [] KAPATA [59] NATHAVITHARANA [7] DORMAN [] BWANGA [] NIEHAUS [7] HUANG [55] MIOTTO [69] CABIBBE [5] AL-MUTAIRI [] HUYEN [56] TOLANI [9] EVANS [] HILLEMANN [5] ALBERT [] ANEK-VORAPONG [5] FENRO [8] KUMAR [6] HILLEMANN [5] LEUTKEMEYER [6] CAUSSE [7] MIOTTO [69] NATHAVITHARANA [7] TUKVADZE [9] RAVEENDRAN [78] LACOMA [6] N GUESSAN [7] ASENCIOS [7] YADAV [9] TOLANI [9] BARNARD [] AURIN [9] ANEK-VORAPONG [5] ASANTE POKU [6] ASENCIOS [7] AUNG [8] CAUSSE [7] ELISEEV [] GAUTHIER [5] HUANG [5] MACEDO [66] NWOFOR [75] SANGSAYUNH [8] SCOTT [8] TESSEMA [88] MTBDRplus V NATHAVITHARANA [7] NATHAVITHARANA [7] Nipro NATHAVITHARANA [7] NATHAVITHARANA [7] 59.6 ( ) 6. ( ) 69.6 ( ) 7.5 ( ) 75. (.8 9.5) 77.8 ( ) ( ) 8. ( ) 8.5 ( ) 8. ( ) 8.( ) 8.6 ( ) 85.7 (. 99.6) 85.9 ( ) ( ) 87.5 ( ) ( ) 87.8 ( ) ( ) ( ) 9.6 ( ) 9. (8. 97.) 9. ( ) 9. ( ) 9.68 ( ) 9. ( ) 9.55 ( ) 9. ( ) 9.7 ( ) 9.7 ( ) 9.7 ( ) 95.6 ( ) 96.5 ( ) 96. ( ) ( ) ( ) 97.6 ( ) 98. ( ) 98.8 ( ) 99.7 ( ). (7.8.). (6.6.). (8.89.). (89..). (9.76.). (9.79.). (76.8.). (79..). (8.56.). (5.7.). (8.5.). (.5.). (75.9.) 8. ( ) 96.7 ( ) 8.97 (78. 9.) 9.7 ( ). (9..) 99. ( ) ( ) 98.5 ( ). (9.89.). (95.89.). (88..). (97..) 98. ( ) 99.6 ( ) 98.6 ( ) 99.5 ( ). (85.75.) ( ). (96.8.) 5. ( ). (86.77.). (9.78.). (9.6.) 9.9 ( ) ( ). (9.89.) 96. ( ). (97.59.) ( ). (95.65.). (9.9.). (98.97.) 95.5 ( ) 8. ( ) ( ) 98.8 ( ) ( ). (8.89.) 9.6 ( ) ( ). (7.9.) 57. (8. 9.). (98.99.) ( ). (9..). (96.55.). (95..). (97.66.). (76.8.) 98.8 ( ). (98..). (98.6.). (9.9.). (96..) 87.5 ( ). (75.9.). (98.5.) 98.6 ( ) 98.6 ( ( ) ( ) Sensitivity % Specificity % FIGURE 6 Forest plots demonstrating sensitivity and specificity of all line probe assays for multidrug-resistant tuberculosis detection for both specimen types compared with phenotypic drug susceptibility testing. TP: true positive; FP: false positive; FN: false negative; TN: true negative. 96.9% respectively for RIF, 89.% and 9.% for INH). Specificity was slightly increased for indirect testing (99.% compared with 98.% for RIF, 99.7% compared with 98.% for INH). The summary point estimates approach the upper left-hand corner of the plots, suggesting good accuracy of LPAs for detection of RIF and INH resistance whether tested directly or indirectly. No studies performed LPA testing on specimens and culture isolates from the same patients precluding direct within-study comparisons. Diagnosis of pulmonary M. tuberculosis using a culture-based reference standard Data to answer this question were limited, as the majority of LPA studies identified by our search criteria did not report results for M. tuberculosis detection. Of the datasets that did report data on M. tuberculosis detection, 5 studies were excluded because they either tested patients who were on treatment or did not specify that patients on treatment were excluded. Six datasets were included in the bivariate analysis [5,, 7, 9, 6, 8], with a total of 5 samples that included 77 (7%) confirmed M. tuberculosis cases tested directly with LPA. Meta-analysis of datasets that reported both sensitivity and specificity revealed a pooled sensitivity of 85.% (95% CI 7. 9.%) and specificity of 98.% (95% CI %) independent of smear-status (table and figure d). Of note, a post hoc bivariate analysis of the datasets (including those that did not exclude patients on treatment) revealed a sensitivity of 9.8% (95% CI %) and specificity of 95.7% (95% CI %). 6